feat(verify): add analyzeO2/Res and decodeMono24 differential tests

Assisted-by: Qwen 3.8 Max Preview

asm/evex.go

@@ -377,6 +377,37 @@ var evexTable = map[string]evexSpec{
 	"VPMOVW2M": {2, 0x29, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
 	"VPMOVD2M": {2, 0x39, 0, 2, -1, vexRM, [3]int{16, 32, 64}},
 	"VPMOVQ2M": {2, 0x39, 1, 2, -1, vexRM, [3]int{16, 32, 64}},
+
+	// EVEX — scalar conversions between vector and general-purpose
+	// registers.  Vector to GPR (two operands: vec/mem source, GPR
+	// destination, vvvv unused): the signed and truncated pair, and the
+	// unsigned forms (EVEX only).
+	"VCVTSD2SI":    {1, 0x2D, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSD2SIQ":   {1, 0x2D, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSS2SI":    {1, 0x2D, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTSS2SIQ":   {1, 0x2D, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSD2SI":   {1, 0x2C, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSD2SIQ":  {1, 0x2C, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSS2SI":   {1, 0x2C, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSS2SIQ":  {1, 0x2C, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTSD2USIL":  {1, 0x79, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSD2USIQ":  {1, 0x79, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTSS2USIL":  {1, 0x79, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTSS2USIQ":  {1, 0x79, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSD2USIL": {1, 0x78, 0, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSD2USIQ": {1, 0x78, 1, 3, -1, vexRM, [3]int{8, 8, 8}},
+	"VCVTTSS2USIL": {1, 0x78, 0, 2, -1, vexRM, [3]int{4, 4, 4}},
+	"VCVTTSS2USIQ": {1, 0x78, 1, 2, -1, vexRM, [3]int{4, 4, 4}},
+	// GPR to vector (three operands: GPR/mem source in r/m, the preserved
+	// vector source in vvvv, vector destination in reg).
+	"VCVTSI2SDL":  {1, 0x2A, 0, 3, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTSI2SDQ":  {1, 0x2A, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VCVTSI2SSL":  {1, 0x2A, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTSI2SSQ":  {1, 0x2A, 1, 2, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VCVTUSI2SDL": {1, 0x7B, 0, 3, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTUSI2SDQ": {1, 0x7B, 1, 3, -1, vexNDS3, [3]int{8, 8, 8}},
+	"VCVTUSI2SSL": {1, 0x7B, 0, 2, -1, vexNDS3, [3]int{4, 4, 4}},
+	"VCVTUSI2SSQ": {1, 0x7B, 1, 2, -1, vexNDS3, [3]int{8, 8, 8}},
 	// EVEX.128/256/512.66.0F38.W0 — sign-extend dwords to qwords; the memory
 	// operand is the narrow source, so disp8×N follows its size (8/16/32 for
 	// the xmm/ymm/zmm destination lengths).
@@ -615,6 +646,12 @@ var evexRound = map[string]bool{
 	"VCVTPD2PS": true, "VCVTPD2UDQ": true, "VCVTTPD2UDQ": true, "VCVTTPD2UQQ": true,
 	"VCVTPS2UDQ": true, "VCVTTPS2UDQ": true, "VCVTPS2UQQ": true, "VCVTTPS2UQQ": true,
 	"VCVTTPD2QQ": true, "VCVTTPS2QQ": true, "VCVTUQQ2PD": true, "VCVTUQQ2PS": true,
+	"VCVTSD2SI": true, "VCVTSD2SIQ": true, "VCVTSS2SI": true, "VCVTSS2SIQ": true,
+	"VCVTSD2USIL": true, "VCVTSD2USIQ": true, "VCVTSS2USIL": true, "VCVTSS2USIQ": true,
+	"VCVTTSD2SI": true, "VCVTTSD2SIQ": true, "VCVTTSS2SI": true, "VCVTTSS2SIQ": true,
+	"VCVTTSD2USIL": true, "VCVTTSD2USIQ": true, "VCVTTSS2USIL": true, "VCVTTSS2USIQ": true,
+	"VCVTSI2SDQ": true, "VCVTSI2SSL": true, "VCVTSI2SSQ": true,
+	"VCVTUSI2SDQ": true, "VCVTUSI2SSL": true, "VCVTUSI2SSQ": true,
 }
 
 // evexBcstN maps an instruction accepting .BCST to the broadcast element
@@ -801,19 +838,23 @@ func (e *enc) encodeEvexNDS3(spec evexSpec, ops []Operand, mask int, sfx evexSuf
 
 // encodeEvexRM encodes the two-operand form: OP src, dst (reg=dst, rm=src,
 // no vvvv), e.g. VCVTQQ2PD.  The destination may be an opmask register (the
-// *2M mask conversions), in which case the vector length comes from the
+// *2M mask conversions) or a general-purpose register (the scalar
-// source.
+// vector-to-GPR conversions); in both cases the vector length comes from
+// the source.
 func (e *enc) encodeEvexRM(spec evexSpec, ops []Operand, mask int, sfx evexSuffix) error {
 	if len(ops) != 2 {
 		return fmt.Errorf("EVEX two-operand instruction expects 2 operands, got %d", len(ops))
 	}
 	src, dst := ops[0], ops[1]
 	dstReg, ok := dst.(Reg)
-	if !ok || (!dstReg.isVec() && !dstReg.mask) {
+	if !ok {
-		return fmt.Errorf("EVEX destination must be a vector or mask register")
+		return fmt.Errorf("EVEX destination must be a register")
 	}
 	ll := dstReg.vecLenBit()
-	if dstReg.mask {
+	if !dstReg.isVec() {
+		// Mask or GPR destination: the length follows the vector source
+		// (128 for a memory source).
+		ll = 0
 		if r, ok := src.(Reg); ok && r.isVec() {
 			ll = r.vecLenBit()
 		}

asm/evex_test.go

@@ -467,6 +467,74 @@ func TestEvexHelperGroundTruth(t *testing.T) {
 	}
 }
 
+// TestEvexGprGroundTruth covers the scalar conversions between vector and
+// general-purpose registers — the signed and truncated VCVT{,T}S{D,S}2SI
+// forms (VEX and EVEX), the unsigned EVEX-only forms, and the GPR-to-vector
+// VCVTSI2*/VCVTUSI2* forms with the preserved vector source in vvvv — byte
+// for byte against the Go assembler, including memory sources and extended
+// GPRs.
+func TestEvexGprGroundTruth(t *testing.T) {
+	mem := func(b Reg) Operand { return Ptr(b, 0, 8) }
+	cases := []struct {
+		name string
+		mnem string
+		ops  []Operand
+		want string
+	}{
+		{"VCVTSD2SI", "VCVTSD2SI", []Operand{vreg(t, "X1"), AX}, "c5fb2dc1"},
+		{"VCVTSD2SIQ", "VCVTSD2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fb2dc1"},
+		{"VCVTSS2SI", "VCVTSS2SI", []Operand{vreg(t, "X1"), AX}, "c5fa2dc1"},
+		{"VCVTSS2SIQ", "VCVTSS2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fa2dc1"},
+		{"VCVTTSD2SI", "VCVTTSD2SI", []Operand{vreg(t, "X1"), AX}, "c5fb2cc1"},
+		{"VCVTTSD2SIQ", "VCVTTSD2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fb2cc1"},
+		{"VCVTTSS2SI", "VCVTTSS2SI", []Operand{vreg(t, "X1"), AX}, "c5fa2cc1"},
+		{"VCVTTSS2SIQ", "VCVTTSS2SIQ", []Operand{vreg(t, "X1"), AX}, "c4e1fa2cc1"},
+		{"VCVTSD2USIL", "VCVTSD2USIL", []Operand{vreg(t, "X1"), AX}, "62f17f0879c1"},
+		{"VCVTSD2USIQ", "VCVTSD2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1ff0879c1"},
+		{"VCVTSS2USIL", "VCVTSS2USIL", []Operand{vreg(t, "X1"), AX}, "62f17e0879c1"},
+		{"VCVTSS2USIQ", "VCVTSS2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1fe0879c1"},
+		{"VCVTTSD2USIL", "VCVTTSD2USIL", []Operand{vreg(t, "X1"), AX}, "62f17f0878c1"},
+		{"VCVTTSD2USIQ", "VCVTTSD2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1ff0878c1"},
+		{"VCVTTSS2USIL", "VCVTTSS2USIL", []Operand{vreg(t, "X1"), AX}, "62f17e0878c1"},
+		{"VCVTTSS2USIQ", "VCVTTSS2USIQ", []Operand{vreg(t, "X1"), AX}, "62f1fe0878c1"},
+		{"VCVTSI2SDL", "VCVTSI2SDL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c5f32ad0"},
+		{"VCVTSI2SDQ", "VCVTSI2SDQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c4e1f32ad0"},
+		{"VCVTSI2SSL", "VCVTSI2SSL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c5f22ad0"},
+		{"VCVTSI2SSQ", "VCVTSI2SSQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "c4e1f22ad0"},
+		{"VCVTUSI2SDL", "VCVTUSI2SDL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f177087bd0"},
+		{"VCVTUSI2SDQ", "VCVTUSI2SDQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f1f7087bd0"},
+		{"VCVTUSI2SSL", "VCVTUSI2SSL", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f176087bd0"},
+		{"VCVTUSI2SSQ", "VCVTUSI2SSQ", []Operand{AX, vreg(t, "X1"), vreg(t, "X2")}, "62f1f6087bd0"},
+		{"VCVTSD2SI mem", "VCVTSD2SI", []Operand{mem(AX), BX}, "c5fb2d18"},
+		{"VCVTSI2SDQ mem", "VCVTSI2SDQ", []Operand{mem(BX), vreg(t, "X1"), vreg(t, "X2")}, "c4e1f32a13"},
+		{"VCVTSD2SIQ hi gpr", "VCVTSD2SIQ", []Operand{vreg(t, "X1"), vreg(t, "R9")}, "c461fb2dc9"},
+		{"VCVTSI2SDQ hi gpr", "VCVTSI2SDQ", []Operand{vreg(t, "R10"), vreg(t, "X1"), vreg(t, "X2")}, "c4c1f32ad2"},
+	}
+	for _, c := range cases {
+		code, err := Encode(c.mnem, c.ops...)
+		if err != nil {
+			t.Errorf("%s: Encode: %v", c.name, err)
+			continue
+		}
+		if got := hexCompact(code); got != c.want {
+			t.Errorf("%s: bytes %s, want %s", c.name, got, c.want)
+			continue
+		}
+		inst, err := x86asm.Decode(code, 64)
+		if err != nil {
+			t.Errorf("%s: Decode(%x): %v", c.name, code, err)
+			continue
+		}
+		// The decoder does not distinguish the Plan 9 SIQ spelling (the
+		// 64-bit GPR destination) from the base name; the W bit carries it.
+		want := c.mnem
+		got := inst.Op.String()
+		if got != want && !(len(want) > len(got) && want[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s", c.name, got)
+		}
+	}
+}
+
 // TestEvexConversionGroundTruth covers the unsigned and truncating VCVT*
 // conversions, the remaining sign/zero-extending moves, the signed/unsigned
 // narrowing stores and the mask/vector conversions, byte for byte against

asm/vex.go

@@ -210,6 +210,24 @@ var vexTable = map[string]vexSpec{
 	"VCVTPD2PSX": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
 	"VCVTPD2PSY": {1, 0x5A, 0, 1, -1, vexRMSrcLen},
 
+	// VEX scalar conversions between vector and general-purpose registers.
+	// Vector to GPR (two operands: vec/mem source, GPR destination, vvvv
+	// unused; the length follows the source).
+	"VCVTSD2SI":   {1, 0x2D, 0, 3, -1, vexRM},
+	"VCVTSD2SIQ":  {1, 0x2D, 1, 3, -1, vexRM},
+	"VCVTSS2SI":   {1, 0x2D, 0, 2, -1, vexRM},
+	"VCVTSS2SIQ":  {1, 0x2D, 1, 2, -1, vexRM},
+	"VCVTTSD2SI":  {1, 0x2C, 0, 3, -1, vexRM},
+	"VCVTTSD2SIQ": {1, 0x2C, 1, 3, -1, vexRM},
+	"VCVTTSS2SI":  {1, 0x2C, 0, 2, -1, vexRM},
+	"VCVTTSS2SIQ": {1, 0x2C, 1, 2, -1, vexRM},
+	// GPR to vector (three operands: GPR/mem source in r/m, the preserved
+	// vector source in vvvv, vector destination in reg).
+	"VCVTSI2SDL": {1, 0x2A, 0, 3, -1, vexNDS3},
+	"VCVTSI2SDQ": {1, 0x2A, 1, 3, -1, vexNDS3},
+	"VCVTSI2SSL": {1, 0x2A, 0, 2, -1, vexNDS3},
+	"VCVTSI2SSQ": {1, 0x2A, 1, 2, -1, vexNDS3},
+
 	// VEX.128/256.66.0F.WIG — word shifts (opdigit selects the shift).
 	"VPSRLW": {1, 0x71, 0, 1, 2, vexShiftImm},
 	"VPSRAW": {1, 0x71, 0, 1, 4, vexShiftImm},

asm/vex_test.go

@@ -39,11 +39,14 @@ func TestVexNDS3(t *testing.T) {
 		}
 		inst, err := x86asm.Decode(code, 64)
 		if err != nil {
-			t.Errorf("%s: Decode(% x): %v", mnem, code, err)
+			t.Errorf("%s: Decode(% x): %v", mnem, err, code)
 			continue
 		}
-		if inst.Op.String() != mnem {
+		// The decoder folds the Plan 9 L/Q GPR-width spellings (VCVTSI2SDL/
-			t.Errorf("%s: decoded as %s (% x)", mnem, inst.Op.String(), code)
+		// SDQ, SSL/SSQ) onto the base name; the W bit carries the width.
+		got := inst.Op.String()
+		if got != mnem && !(len(mnem) > len(got) && mnem[:len(got)] == got) {
+			t.Errorf("%s: decoded as %s (% x)", mnem, got, code)
 		}
 	}
 }

cmd/gasm/main.go

@@ -24,11 +24,12 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/lint"
 	"sourcedock.dev/petrbalvin/gasm-devkit/lsp"
 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
 )
 
 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.15.0"
+var version = "0.23.0"
 
 func main() {
 	if len(os.Args) < 2 {
@@ -46,6 +47,8 @@ func main() {
 		os.Exit(cmdLint(os.Args[2:]))
 	case "asm":
 		os.Exit(cmdAsm(os.Args[2:]))
+	case "verify":
+		os.Exit(cmdVerify(os.Args[2:]))
 	case "lsp":
 		os.Exit(cmdLSP(os.Args[2:]))
 	case "version", "--version", "-V":
@@ -80,6 +83,7 @@ Commands:
   fmt       canonicalise formatting (gofmt for assembly)
   lint      run static checks
   asm       assemble .s files to machine code (amd64)
+  verify    JIT-assemble and run dynamic checks (amd64)
   lsp       run the language server over stdio
   version   print the version (same as --version)
 
@@ -466,3 +470,88 @@ requires -p, the package path, and the installed Go toolchain).
 	}
 	return 0
 }
+
+func cmdVerify(args []string) int {
+	fs := newCommand("verify", "gasm verify [-smoke] [-abi] [-profile] <file.s>", `
+Assemble FILE (amd64), map it into executable memory and report the available
+functions.  This confirms the assembled image is self-consistent (no
+unresolved external symbols) and executable — the prerequisite for dynamic
+testing.
+
+With -smoke, each NOSPLIT function is called with a zeroed argument block to
+confirm the JIT trampoline works end-to-end.  This is safe only for functions
+that tolerate nil pointers and zero lengths in their arguments.
+
+With -abi, each function is called with sentinel values in the callee-saved
+registers (BP, R14) and a red-zone canary below SP; violations are reported.
+
+With -profile, the static basic-block structure is listed for each function.
+`)
+	smoke := fs.Bool("smoke", false, "call each NOSPLIT function with zeroed args")
+	abi := fs.Bool("abi", false, "run ABI-checking calls (sentinel registers + red zone)")
+	profile := fs.Bool("profile", false, "list basic-block structure per function")
+	fs.Parse(args)
+	if fs.NArg() != 1 {
+		fmt.Fprintln(os.Stderr, "usage: gasm verify [-smoke] [-abi] [-profile] <file.s>")
+		return 2
+	}
+	path := fs.Arg(0)
+	if arch.FromFilename(path) != arch.AMD64 {
+		fmt.Fprintln(os.Stderr, "gasm verify: only amd64 is supported")
+		return 1
+	}
+
+	k, err := verify.Load(path)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm verify: %v\n", err)
+		return 1
+	}
+	defer k.Close()
+
+	names := k.FuncNames()
+	fmt.Printf("%s: %d functions JIT-loaded\n", path, len(names))
+	rc := 0
+	for _, name := range names {
+		fl, _ := k.Func(name)
+		flags := ""
+		if fl.NoSplit {
+			flags = " NOSPLIT"
+		}
+		fmt.Printf("  %s: %d bytes, args=%d, frame=%d%s\n", name, fl.Size, fl.Args, fl.Frame, flags)
+
+		if *profile {
+			blocks, err := k.Blocks(name)
+			if err != nil {
+				fmt.Printf("    profile: %v\n", err)
+			} else {
+				fmt.Printf("    blocks: %d\n", len(blocks))
+			}
+		}
+
+		if *smoke && fl.NoSplit {
+			args := make([]byte, fl.Args)
+			_, err := k.CallFunc(name, args)
+			if err != nil {
+				fmt.Printf("    smoke: FAIL — %v\n", err)
+				rc = 1
+			} else {
+				fmt.Printf("    smoke: OK\n")
+			}
+		}
+
+		if *abi && fl.NoSplit {
+			args := make([]byte, fl.Args)
+			_, report, err := k.CallFuncChecked(name, args)
+			if err != nil {
+				fmt.Printf("    abi: FAIL — %v\n", err)
+				rc = 1
+			} else if !report.OK() {
+				fmt.Printf("    abi: %s\n", report)
+				rc = 1
+			} else {
+				fmt.Printf("    abi: clean\n")
+			}
+		}
+	}
+	return rc
+}

docs/ARCHITECTURE.md

@@ -242,7 +242,9 @@ helper and conversion tail (VRCP14*, VRSQRT14*, VGETEXP*, VGETMANT*,
 VSCALEF*, VRNDSCALE*, VREDUCE*, VFIXUPIMM*, VRANGE*, VFPCLASS* with an
 opmask destination, and the VCVT* conversions — signed, unsigned and
 truncating, including the length-suffixed X/Y spellings and the
-mask/vector conversions VPMOVM2*/VPMOV*2M), and gather/scatter with VSIB addressing — both the
+mask/vector conversions VPMOVM2*/VPMOV*2M, and the scalar conversions
+between vector and general-purpose registers (VCVT{,T}S{D,S}2SI{,Q} and
+the unsigned forms, VCVTSI2*/VCVTUSI2*), and gather/scatter with VSIB addressing — both the
 VEX spelling with a vector mask register and the EVEX spelling with an
 explicit K mask, where the EVEX length follows the VSIB index register,
 not the data register.  The EVEX mnemonic
@@ -282,6 +284,31 @@ references and the implicit funcdata/DWARF symbols remain future work (the
 linker fills the latter's defaults); the rest of Phase 2 is those, the
 remaining EVEX forms and the other architectures.
 
+### `verify`
+
+The dynamic-analysis substrate (Phase 3).  It JIT-loads assembled images into
+executable memory and invokes them directly, enabling differential testing,
+runtime ABI checks and coverage profiling.
+
+The execution model is pure Go (stdlib only).  `Map` copies machine code into
+an anonymous `syscall.Mmap` mapping and enforces W^X (write the bytes, then
+`mprotect` to read-execute).  `Call` prepares a stack whose first word is the
+address of an assembly trampoline (`leaveJIT`), lays the ABI0 argument
+block after it, switches to that stack via `enterJIT` (which saves the Go
+stack pointer in a package global and jumps to the target), and recovers
+control when the function RETs into `leaveJIT` (which restores the Go stack
+and returns).  A 64-byte pad below the return address accommodates the
+ABIInternal wrapper that the Go runtime interposes on assembly functions.
+
+`Load` / `LoadSource` / `LoadAST` parse, assemble and map a `.s` file in one
+step, returning a `Kernel` whose `CallFunc` method marshals the argument block
+by name.  The image must be self-contained (no external relocations); the
+assembler’s `Image.Bytes()` provides the code-and-data concatenation.
+
+The `gasm verify` CLI subcommand exposes this: it loads a file, reports the
+available functions and (with `-smoke`) calls each NOSPLIT function with zeroed
+arguments to confirm the trampoline round-trips.
+
 ## Extension points
 
 - **New architecture:** add an entry to the generator in `_gen`, run

docs/DEFERRED.md

@@ -0,0 +1,56 @@
+# Deferred decisions
+
+Design decisions deliberately postponed, with enough context to pick them up
+again without re-deriving the analysis.  Each entry records what is deferred,
+why, the options on the table, and the trigger that should reopen it.
+
+---
+
+## GOOBJ external (cross-package) symbol references
+
+**Status:** deferred (v0.15.0, 2026-08-02).  The GOOBJ emitter resolves only
+symbols defined in the file being assembled; a reference to any other symbol
+is rejected.
+
+**Why it is deferred.**  GOOBJ symbol references are *positional*: a
+reference is a `{PkgIdx, SymIdx}` pair, where `SymIdx` is the index of the
+symbol in the *referenced package's* symbol-definition table.  That ordering
+is not derivable from the reference site — it lives in the referenced
+package's gc export data (the iexport binary format, which evolves with the
+toolchain).  `cmd/asm` reads it with `cmd/internal` readers gasm cannot
+import, so emitting external references means either parsing export data
+ourselves or taking a dependency that does.
+
+**What works today.**  Single-package objects: every symbol the file defines
+(as `TEXT` or `GLOBL`, static or exported) and every reference to them.
+This covers the production use case — the go-flac / go-lz4 kernels carry no
+`FUNCDATA`/`PCDATA`, hence no references into `runtime`, and the Go side
+references the assembly symbols, never the reverse.  Such a package builds
+with its assembly object replaced by a gasm-emitted one.
+
+**The options, when we return.**
+
+1. **`golang.org/x/tools/go/gcexportdata` as a production dependency.**
+   The straightforward path: read each imported package's export file
+   (paths from `-importcfg` or `go list -export`), assign symbol indices in
+   its symbol order, write `PkgIndex`/`Autolib` entries (fingerprints from
+   the export files' build IDs) and positional references.  Robust across
+   toolchain versions — `x/tools` tracks the format.  **Cost:** the first
+   production dependency beyond the standard library, an explicit deviation
+   from the "production code depends only on the standard library"
+   principle in the README.  Requires the user's explicit agreement.
+2. **A minimal iexport parser of our own.**  Preserves self-containment.
+   Substantial effort and inherently fragile: the format is an internal
+   contract that changes with Go releases, so the parser needs a
+   version-gated fallback and regression tests against several toolchains.
+3. **Shell out to the toolchain for symbol metadata.**  Consistent with the
+   existing GOOBJ preamble probe (which already runs `go tool asm`), but no
+   toolchain command exposes a package's symbols *in definition-index
+   order* — `go tool nm` sorts differently — so this does not solve the
+   core problem on its own; it would only feed option 1 or 2.
+
+**Trigger to reopen.**  An assembly file that needs a cross-package
+reference — in practice `FUNCDATA $…, runtime·…(SB)` (stack maps / GC
+metadata written in assembly), or any kernel that calls into another
+package directly.  Until then, option 3's limitation is moot and the
+single-package emitter suffices.

justfile

@@ -3,7 +3,7 @@
 
 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).
 
-version := "0.15.0"
+version := "0.23.0"
 
 default:
     @just --list

testdata/verify/abi_amd64.s

@@ -0,0 +1,28 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// func cleanAdd(a, b int64) int64
+// A well-behaved function that preserves all callee-saved registers.
+TEXT ·cleanAdd(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+// func dirtyBP(a int64) int64
+// Deliberately clobbers BP (an ABI violation for a NOSPLIT frame=0 function).
+TEXT ·dirtyBP(SB), NOSPLIT, $0-16
+	MOVQ $0x1234, BP
+	MOVQ a+0(FP), AX
+	MOVQ AX, ret+8(FP)
+	RET
+
+// func dirtyR14(a int64) int64
+// Deliberately clobbers R14 (the goroutine pointer — a serious ABI violation).
+TEXT ·dirtyR14(SB), NOSPLIT, $0-16
+	MOVQ $0x5678, R14
+	MOVQ a+0(FP), AX
+	MOVQ AX, ret+8(FP)
+	RET

testdata/verify/basic_amd64.s

@@ -0,0 +1,67 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// func add(a, b int64) int64
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+// func sum(data []int64) int64
+// Sums all elements of the slice.
+TEXT ·sum(SB), NOSPLIT, $0-32
+	MOVQ data_base+0(FP), SI
+	MOVQ data_len+8(FP), CX
+	XORQ AX, AX
+	TESTQ CX, CX
+	JZ   sum_done
+
+sum_loop:
+	ADDQ (SI), AX
+	ADDQ $8, SI
+	DECQ CX
+	JNZ  sum_loop
+
+sum_done:
+	MOVQ AX, ret+24(FP)
+	RET
+
+// func wideCopy(dst, src []byte)
+// Non-overlapping copy of min(len(dst), len(src)) bytes using 32-byte moves.
+TEXT ·wideCopy(SB), NOSPLIT, $0-48
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R8
+	CMPQ BX, R8
+	JLE  wc_have_n
+	MOVQ R8, BX
+
+wc_have_n:
+	CMPQ BX, $32
+	JB   wc_small
+
+	VMOVDQU (SI), Y0
+	VMOVDQU Y0, (DI)
+	VMOVDQU -32(SI)(BX*1), Y0
+	VMOVDQU Y0, -32(DI)(BX*1)
+	VZEROUPPER
+	RET
+
+wc_small:
+	TESTQ BX, BX
+	JZ    wc_done
+
+wc_byte:
+	MOVB (SI), R8B
+	MOVB R8B, (DI)
+	INCQ SI
+	INCQ DI
+	DECQ BX
+	JNZ  wc_byte
+
+wc_done:
+	RET

verify/abi_amd64.go

@@ -0,0 +1,130 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build amd64
+
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+// abiResult records register-clobber violations detected by the ABI-checking
+// trampoline.  Bit 0: BP clobbered.  Bit 1: R14 clobbered.
+var abiResult uint64
+
+// savedBP holds the caller's frame pointer across the ABI-checked JIT call.
+// Referenced by enterJITChecked to satisfy go vet's save-before-clobber rule.
+var savedBP uintptr
+
+// leaveCheckedPtr is initialised by the linker from the GLOBL/DATA in
+// abi_amd64.s: it holds the raw address of leaveJITCheckedRaw (which has
+// no ABIInternal wrapper, so the JIT function RETs directly into it).
+var leaveCheckedPtr uintptr
+
+// enterJITChecked sets sentinels in BP and R14, switches to the prepared
+// stack and jumps to fn.
+//
+//go:nosplit
+func enterJITChecked(fn uintptr, stack uintptr)
+
+// leaveJITCheckedRaw is the raw return trampoline for ABI checks.  Its
+// address is obtained from the GLOBL in abi_amd64.s (leaveCheckedPtr),
+// which points to the .abi0 code — NOT the ABIInternal wrapper that this
+// declaration would generate.  The declaration exists solely to satisfy
+// go vet's "missing Go declaration" check.
+//
+//go:nosplit
+func leaveJITCheckedRaw()
+
+// ABIReport describes the result of an ABI-checking call.
+type ABIReport struct {
+	BPClobbered  bool // BP was modified by the function
+	R14Clobbered bool // R14 (goroutine pointer) was modified
+	RedZoneHit   bool // the 128-byte red zone below SP was written
+}
+
+// OK returns true when no violations were detected.
+func (r ABIReport) OK() bool {
+	return !r.BPClobbered && !r.R14Clobbered && !r.RedZoneHit
+}
+
+// String returns a human-readable summary.
+func (r ABIReport) String() string {
+	if r.OK() {
+		return "ABI clean"
+	}
+	s := "ABI violation:"
+	if r.BPClobbered {
+		s += " BP clobbered"
+	}
+	if r.R14Clobbered {
+		s += " R14 clobbered"
+	}
+	if r.RedZoneHit {
+		s += " red-zone written"
+	}
+	return s
+}
+
+// redZoneSize is the System V AMD64 red zone: 128 bytes below SP that a
+// leaf function may use without adjusting SP.  Go does not use the red zone,
+// so any write there is a bug.
+const redZoneSize = 128
+
+// redZoneFill is the byte pattern used to detect red-zone writes.
+const redZoneFill = 0xA5
+
+// CallChecked invokes the function with ABI sentinels and a red-zone
+// canary, returning both the argument block (with results) and an ABIReport.
+func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
+	report := ABIReport{}
+
+	// Reset the global result.
+	abiResult = 0
+
+	// Prepare the stack: [red-zone canary][padding][leaveJITCheckedRaw][args...]
+	// The red zone sits below the initial SP, so the function would have to
+	// write below SP to corrupt it.
+	totalSize := redZoneSize + stackPad + 8 + len(args) + 64
+	stackMem, err := syscall.Mmap(-1, 0, totalSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, report, fmt.Errorf("verify: stack mmap: %w", err)
+	}
+	defer syscall.Munmap(stackMem)
+
+	// Fill the red zone with the canary pattern.
+	for i := 0; i < redZoneSize; i++ {
+		stackMem[i] = redZoneFill
+	}
+
+	// Return address and args after the red zone and padding.
+	retOff := redZoneSize + stackPad
+	binary.LittleEndian.PutUint64(stackMem[retOff:retOff+8], uint64(leaveCheckedPtr))
+	copy(stackMem[retOff+8:], args)
+
+	stackBase := uintptr(unsafe.Pointer(&stackMem[retOff]))
+	enterJITChecked(fnAddr, stackBase)
+
+	// Read the register-clobber result.
+	res := abiResult
+	report.BPClobbered = res&1 != 0
+	report.R14Clobbered = res&2 != 0
+
+	// Check the red zone.
+	for i := 0; i < redZoneSize; i++ {
+		if stackMem[i] != redZoneFill {
+			report.RedZoneHit = true
+			break
+		}
+	}
+
+	// Copy out the argument area.
+	out := make([]byte, len(args))
+	copy(out, stackMem[retOff+8:retOff+8+len(args)])
+	return out, report, nil
+}

verify/abi_amd64.s

@@ -0,0 +1,61 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// ABI-checking trampoline.  Sets sentinel values in the callee-saved
+// registers (BP, R14) before entering the JIT function and checks whether
+// they survived on return.
+//
+// The return trampoline (leaveJITCheckedRaw) is a raw TEXT symbol with no
+// Go function declaration, so the toolchain does NOT interpose an
+// ABIInternal wrapper — the JIT function RETs directly into the check code,
+// which sees the registers exactly as the function left them.
+//
+// Go ABI0 on amd64 guarantees:
+//   - BP is callee-saved (NOSPLIT frame=0 functions must not touch it).
+//   - R14 holds the goroutine pointer and must survive across any call.
+
+// Sentinel values chosen to be unlikely in normal execution.
+#define SENTINEL_BP  0xDEADBEEFCAFEF00D
+#define SENTINEL_R14 0x0BADF00DDEADBEEF
+
+// GLOBL holding the raw address of the leave trampoline, read by Go.
+GLOBL ·leaveCheckedPtr(SB), NOPTR, $8
+DATA  ·leaveCheckedPtr(SB)/8, $·leaveJITCheckedRaw(SB)
+
+// func enterJITChecked(fn uintptr, stack uintptr)
+// Sets sentinels in BP and R14, switches to the prepared stack and jumps
+// to fn.  The prepared stack's return address must be leaveJITCheckedRaw
+// (read from leaveCheckedPtr).
+TEXT ·enterJITChecked(SB), NOSPLIT, $0-16
+	MOVQ fn+0(FP), AX         // target (before SP switch)
+	MOVQ SP, ·savedSP(SB)     // preserve Go stack
+	MOVQ BP, ·savedBP(SB)     // preserve frame pointer (vet requires save before clobber)
+	MOVQ $SENTINEL_BP, BP     // sentinel in BP
+	MOVQ $SENTINEL_R14, R14   // sentinel in R14
+	MOVQ stack+8(FP), SP      // switch to prepared stack
+	JMP  AX
+
+// leaveJITCheckedRaw is the raw return trampoline.  It has NO Go function
+// declaration, so no ABIInternal wrapper is generated — the JIT function's
+// RET lands here directly, seeing BP and R14 exactly as the function left
+// them.  It checks the sentinels, records violations in abiResult, then
+// restores the Go stack and returns.
+TEXT ·leaveJITCheckedRaw(SB), NOSPLIT, $0-0
+	// Check BP against the sentinel.
+	MOVQ $SENTINEL_BP, CX
+	CMPQ BP, CX
+	JEQ  bp_ok
+	ORQ  $1, ·abiResult(SB)
+
+bp_ok:
+	// Check R14 against the sentinel.
+	MOVQ $SENTINEL_R14, CX
+	CMPQ R14, CX
+	JEQ  r14_ok
+	ORQ  $2, ·abiResult(SB)
+
+r14_ok:
+	MOVQ ·savedSP(SB), SP
+	RET

verify/abi_other.go

@@ -0,0 +1,26 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build !amd64
+
+package verify
+
+import "fmt"
+
+// ABIReport describes the result of an ABI-checking call.
+type ABIReport struct {
+	BPClobbered  bool
+	R14Clobbered bool
+	RedZoneHit   bool
+}
+
+// OK returns true when no violations were detected.
+func (r ABIReport) OK() bool { return false }
+
+// String returns a human-readable summary.
+func (r ABIReport) String() string { return "verify: ABI checks require amd64" }
+
+// CallChecked is unavailable on non-amd64 architectures.
+func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
+	return nil, ABIReport{}, fmt.Errorf("verify: ABI checks require amd64")
+}

verify/abi_test.go

@@ -0,0 +1,145 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"testing"
+	"unsafe"
+)
+
+func loadABIKernel(t *testing.T) *Kernel {
+	t.Helper()
+	k, err := Load("../testdata/verify/abi_amd64.s")
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestABIClean(t *testing.T) {
+	k := loadABIKernel(t)
+
+	args := make([]byte, 24)
+	PutUint64(args, 0, 3)
+	PutUint64(args, 8, 4)
+
+	out, report, err := k.CallFuncChecked("cleanAdd", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked: %v", err)
+	}
+	if got := int64(GetUint64(out, 16)); got != 7 {
+		t.Errorf("cleanAdd(3, 4) = %d, want 7", got)
+	}
+	if !report.OK() {
+		t.Errorf("cleanAdd: %s", report)
+	}
+}
+
+func TestABIBPClobbered(t *testing.T) {
+	k := loadABIKernel(t)
+
+	args := make([]byte, 16)
+	PutUint64(args, 0, 42)
+
+	out, report, err := k.CallFuncChecked("dirtyBP", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked: %v", err)
+	}
+	if got := int64(GetUint64(out, 8)); got != 42 {
+		t.Errorf("dirtyBP(42) = %d, want 42", got)
+	}
+	if !report.BPClobbered {
+		t.Error("dirtyBP: expected BP clobbered, but report says clean")
+	}
+	if report.R14Clobbered {
+		t.Error("dirtyBP: R14 should not be clobbered")
+	}
+}
+
+func TestABIR14Clobbered(t *testing.T) {
+	k := loadABIKernel(t)
+
+	args := make([]byte, 16)
+	PutUint64(args, 0, 99)
+
+	out, report, err := k.CallFuncChecked("dirtyR14", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked: %v", err)
+	}
+	if got := int64(GetUint64(out, 8)); got != 99 {
+		t.Errorf("dirtyR14(99) = %d, want 99", got)
+	}
+	if !report.R14Clobbered {
+		t.Error("dirtyR14: expected R14 clobbered, but report says clean")
+	}
+	if report.BPClobbered {
+		t.Error("dirtyR14: BP should not be clobbered")
+	}
+}
+
+// TestABILZ4Kernels verifies that the production go-lz4 kernels are ABI-clean:
+// they preserve BP and R14 and do not write into the red zone.
+func TestABILZ4Kernels(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	// wideCopyAVX2 with a real copy.
+	src := make([]byte, 128)
+	for i := range src {
+		src[i] = byte(i)
+	}
+	dst := make([]byte, 128)
+
+	args := make([]byte, 48)
+	PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+	PutUint64(args, 8, 128)
+	PutUint64(args, 16, 128)
+	PutPtr(args, 24, unsafe.Pointer(&src[0]))
+	PutUint64(args, 32, 128)
+	PutUint64(args, 40, 128)
+
+	_, report, err := k.CallFuncChecked("wideCopyAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked(wideCopyAVX2): %v", err)
+	}
+	if !report.OK() {
+		t.Errorf("wideCopyAVX2: %s", report)
+	}
+
+	// decodeBlockAVX2 with a simple block.
+	decSrc := []byte{0x50, 'H', 'e', 'l', 'l', 'o'}
+	decDst := make([]byte, 64)
+
+	decArgs := make([]byte, 64)
+	PutPtr(decArgs, 0, unsafe.Pointer(&decSrc[0]))
+	PutUint64(decArgs, 8, uint64(len(decSrc)))
+	PutUint64(decArgs, 16, uint64(cap(decSrc)))
+	PutPtr(decArgs, 24, unsafe.Pointer(&decDst[0]))
+	PutUint64(decArgs, 32, uint64(len(decDst)))
+	PutUint64(decArgs, 40, uint64(cap(decDst)))
+
+	_, report, err = k.CallFuncChecked("decodeBlockAVX2", decArgs)
+	if err != nil {
+		t.Fatalf("CallFuncChecked(decodeBlockAVX2): %v", err)
+	}
+	if !report.OK() {
+		t.Errorf("decodeBlockAVX2: %s", report)
+	}
+}
+
+func TestCallFuncCheckedErrors(t *testing.T) {
+	k := loadABIKernel(t)
+
+	// Nonexistent function.
+	_, _, err := k.CallFuncChecked("nope", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+
+	// Arg block too small.
+	_, _, err = k.CallFuncChecked("cleanAdd", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for too-small arg block")
+	}
+}

verify/avx512_test.go

@@ -0,0 +1,144 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"math/rand"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+const lz4AVX512Path = "../../go-libraries/go-lz4/avx512_amd64.s"
+
+func loadLZ4AVX512Kernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(lz4AVX512Path); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(lz4AVX512Path)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", lz4AVX512Path, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestAVX512DecodeKnownAnswers(t *testing.T) {
+	k := loadLZ4AVX512Kernel(t)
+
+	tests := []struct {
+		name     string
+		src      []byte
+		wantN    int
+		wantCode int
+	}{
+		{"literals_only", []byte{0x50, 'H', 'e', 'l', 'l', 'o'}, 5, 0},
+		{"literals_and_match", []byte{0x54, 'A', 'A', 'A', 'A', 'A', 0x05, 0x00, 0x30, 'B', 'B', 'B'}, 16, 0},
+		{"overlapping", []byte{0x14, 'X', 0x01, 0x00, 0x10, 'Y'}, 10, 0},
+		{"malformed", []byte{0x50, 'H', 'e'}, 0, 1},
+		{"zero_offset", []byte{0x14, 'X', 0x00, 0x00}, 0, 2},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dst := make([]byte, 64)
+			args := make([]byte, 64)
+			PutPtr(args, 0, unsafe.Pointer(&tt.src[0]))
+			PutUint64(args, 8, uint64(len(tt.src)))
+			PutUint64(args, 16, uint64(cap(tt.src)))
+			PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+			PutUint64(args, 32, uint64(len(dst)))
+			PutUint64(args, 40, uint64(cap(dst)))
+
+			out, err := k.CallFunc("decodeBlockAVX512", args)
+			if err != nil {
+				t.Fatalf("CallFunc: %v", err)
+			}
+			n := int(GetUint64(out, 48))
+			code := int(GetUint64(out, 56))
+			if n != tt.wantN || code != tt.wantCode {
+				t.Errorf("got (n=%d, code=%d), want (n=%d, code=%d)", n, code, tt.wantN, tt.wantCode)
+			}
+		})
+	}
+}
+
+func TestAVX512DifferentialFuzz(t *testing.T) {
+	k := loadLZ4AVX512Kernel(t)
+	rng := rand.New(rand.NewSource(77))
+
+	for i := 0; i < 3000; i++ {
+		wantSize := 1 + rng.Intn(4096)
+		src := genLZ4Block(rng, wantSize)
+		dstSize := wantSize + 64
+
+		goDst := make([]byte, dstSize)
+		goN, goCode := decodeBlockGo(src, goDst)
+
+		jitDst := make([]byte, dstSize)
+		args := make([]byte, 64)
+		if len(src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(len(src)))
+		PutUint64(args, 16, uint64(cap(src)))
+		if dstSize > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 32, uint64(dstSize))
+		PutUint64(args, 40, uint64(cap(jitDst)))
+
+		out, err := k.CallFunc("decodeBlockAVX512", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", i, err)
+		}
+		jitN := int(GetUint64(out, 48))
+		jitCode := int(GetUint64(out, 56))
+
+		if jitCode != goCode {
+			t.Fatalf("iter %d: code mismatch: JIT=%d Go=%d", i, jitCode, goCode)
+		}
+		if jitCode != 0 {
+			continue
+		}
+		if jitN != goN {
+			t.Fatalf("iter %d: n mismatch: JIT=%d Go=%d", i, jitN, goN)
+		}
+		if !bytes.Equal(jitDst[:jitN], goDst[:goN]) {
+			t.Fatalf("iter %d: output mismatch (n=%d)", i, jitN)
+		}
+	}
+}
+
+func TestAVX512WideCopy(t *testing.T) {
+	k := loadLZ4AVX512Kernel(t)
+
+	sizes := []int{0, 1, 31, 32, 63, 64, 65, 127, 128, 256, 1024}
+	for _, n := range sizes {
+		src := make([]byte, n)
+		for i := range src {
+			src[i] = byte(i*11 + 3)
+		}
+		dst := make([]byte, n)
+
+		args := make([]byte, 48)
+		if n > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(n))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(n))
+
+		_, err := k.CallFunc("wideCopyAVX512", args)
+		if err != nil {
+			t.Fatalf("wideCopyAVX512(n=%d): %v", n, err)
+		}
+		if !bytes.Equal(dst, src) {
+			t.Errorf("wideCopyAVX512(n=%d): mismatch", n)
+		}
+	}
+}

verify/call_amd64.go

@@ -0,0 +1,77 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build amd64
+
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"reflect"
+	"syscall"
+	"unsafe"
+)
+
+// savedSP holds the Go stack pointer while a JIT call is in flight.
+// Referenced by the assembly trampoline (trampoline_amd64.s).
+var savedSP uintptr
+
+// enterJIT switches to the prepared stack and jumps to fn.
+// It does not return normally; the JIT function's RET transfers control
+// to leaveJIT, which restores the Go stack.
+//
+//go:nosplit
+func enterJIT(fn uintptr, stack uintptr)
+
+// leaveJIT restores the Go stack after a JIT function returns.
+// Its address is placed as the return address on the prepared stack.
+//
+//go:nosplit
+func leaveJIT()
+
+// leaveJITAddr is the machine address of leaveJIT, resolved once at init.
+var leaveJITAddr uintptr
+
+func init() {
+	leaveJITAddr = reflect.ValueOf(leaveJIT).Pointer()
+}
+
+// stackPad is padding below the return address on the prepared stack.
+// The ABIInternal wrapper that leaveJIT's address resolves to executes
+// PUSHQ BP and CALL before reaching the raw assembly, writing up to 16
+// bytes below the return-address slot.  64 bytes of headroom is ample.
+const stackPad = 64
+
+// Call invokes the assembled function at fnAddr with the given ABI0 argument
+// block (the raw bytes that would appear at FP+0).  It returns the argument
+// block after the call, which contains any results the function wrote back
+// (the ABI0 convention shares the argument area for inputs and outputs).
+//
+// The function must be NOSPLIT (no stack growth) and must not reference
+// external symbols — the image is self-contained.
+func Call(fnAddr uintptr, args []byte) ([]byte, error) {
+	// Prepare the stack: [padding][leaveJIT addr][args...]
+	stackSize := stackPad + 8 + len(args) + 64 // padding + ret + args + safety
+	stackMem, err := syscall.Mmap(-1, 0, stackSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, fmt.Errorf("verify: stack mmap: %w", err)
+	}
+	defer syscall.Munmap(stackMem)
+
+	// The return address sits after the padding; the function's SP will
+	// point here, leaving stackPad bytes below for the wrapper's pushes.
+	retOff := stackPad
+	binary.LittleEndian.PutUint64(stackMem[retOff:retOff+8], uint64(leaveJITAddr))
+	// The ABI0 argument area follows the return address.
+	copy(stackMem[retOff+8:], args)
+
+	stackBase := uintptr(unsafe.Pointer(&stackMem[retOff]))
+	enterJIT(fnAddr, stackBase)
+
+	// Copy out the (possibly modified) argument area.
+	out := make([]byte, len(args))
+	copy(out, stackMem[retOff+8:retOff+8+len(args)])
+	return out, nil
+}

verify/call_other.go

@@ -0,0 +1,13 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build !amd64
+
+package verify
+
+import "fmt"
+
+// Call is unavailable on non-amd64 architectures.
+func Call(fnAddr uintptr, args []byte) ([]byte, error) {
+	return nil, fmt.Errorf("verify: JIT execution requires amd64")
+}

verify/coverage.go

@@ -0,0 +1,103 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"fmt"
+	"sort"
+)
+
+// Block describes one basic block within a function: a maximal sequence of
+// instructions with a single entry point (a label or the function start) and
+// a single exit (a jump, conditional jump or RET).
+type Block struct {
+	Offset int    // byte offset within the function
+	Label  string // label name ("" for the entry block)
+}
+
+// Blocks identifies the basic blocks of a function from its local labels.
+// Each label is a potential jump target and therefore a block boundary; the
+// function entry (offset 0) is always a block.  The blocks are returned in
+// ascending offset order.
+func (k *Kernel) Blocks(name string) ([]Block, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+
+	blocks := []Block{{Offset: 0, Label: "(entry)"}}
+	// Build a reverse map: offset → label name.
+	offToLabel := make(map[int]string, len(fl.Labels))
+	for label, off := range fl.Labels {
+		if off > 0 && off < fl.Size {
+			offToLabel[off] = label
+		}
+	}
+	// Collect and sort offsets.
+	offsets := make([]int, 0, len(offToLabel))
+	for off := range offToLabel {
+		offsets = append(offsets, off)
+	}
+	sort.Ints(offsets)
+	for _, off := range offsets {
+		blocks = append(blocks, Block{Offset: off, Label: offToLabel[off]})
+	}
+	return blocks, nil
+}
+
+// BlockCount returns the number of identified basic blocks for the function.
+func (k *Kernel) BlockCount(name string) (int, error) {
+	blocks, err := k.Blocks(name)
+	if err != nil {
+		return 0, err
+	}
+	return len(blocks), nil
+}
+
+// PathFingerprint is the observable output of one function execution: the
+// values written back into the result slots of the argument block.  Two
+// executions that produce the same fingerprint took observationally
+// equivalent paths (though they may differ internally).
+type PathFingerprint struct {
+	Results []uint64 // the result words from the arg block
+}
+
+// ProfilePaths runs the function with each of the given argument blocks and
+// collects the distinct output fingerprints.  This measures path diversity:
+// how many observationally different execution paths the input corpus
+// exercises.  Combined with Blocks (the static block count), it gives a
+// lower bound on code coverage.
+func (k *Kernel) ProfilePaths(name string, argSets [][]byte, resultOffsets []int) ([]PathFingerprint, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+
+	seen := map[string]bool{}
+	var paths []PathFingerprint
+
+	for _, args := range argSets {
+		if len(args) < fl.Args {
+			return nil, fmt.Errorf("verify: %s: arg block too small", name)
+		}
+		out, err := k.CallFunc(name, args)
+		if err != nil {
+			return nil, err
+		}
+		fp := PathFingerprint{}
+		key := ""
+		for _, off := range resultOffsets {
+			v := GetUint64(out, off)
+			fp.Results = append(fp.Results, v)
+			key += fmt.Sprintf("%016x", v)
+		}
+		if !seen[key] {
+			seen[key] = true
+			paths = append(paths, fp)
+		}
+	}
+	return paths, nil
+}

verify/coverage_test.go

@@ -0,0 +1,85 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"testing"
+	"unsafe"
+)
+
+func TestBlocks(t *testing.T) {
+	k := loadBasic(t)
+
+	// The "sum" function has labels: sum_done, sum_loop.
+	blocks, err := k.Blocks("sum")
+	if err != nil {
+		t.Fatalf("Blocks(sum): %v", err)
+	}
+	if len(blocks) < 3 {
+		t.Errorf("sum: expected at least 3 blocks (entry + 2 labels), got %d", len(blocks))
+	}
+	if blocks[0].Offset != 0 {
+		t.Errorf("first block offset = %d, want 0", blocks[0].Offset)
+	}
+	t.Logf("sum blocks: %v", blocks)
+}
+
+func TestBlockCount(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	n, err := k.BlockCount("decodeBlockAVX2")
+	if err != nil {
+		t.Fatalf("BlockCount: %v", err)
+	}
+	// The decoder has many labels (dec_loop, dec_malformed, etc.).
+	if n < 10 {
+		t.Errorf("decodeBlockAVX2: expected at least 10 blocks, got %d", n)
+	}
+	t.Logf("decodeBlockAVX2: %d basic blocks", n)
+}
+
+func TestProfilePaths(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	// Build a corpus of varied LZ4 blocks.
+	var argSets [][]byte
+	blocks := []struct {
+		src     []byte
+		dstSize int
+	}{
+		{[]byte{0x00}, 16},                          // empty
+		{[]byte{0x50, 'H', 'e', 'l', 'l', 'o'}, 16}, // literals only
+		{[]byte{0x54, 'A', 'A', 'A', 'A', 'A', 5, 0, 0x30, 'B', 'B', 'B'}, 32}, // match
+		{[]byte{0x14, 'X', 1, 0, 0x10, 'Y'}, 16},                               // overlapping
+		{[]byte{0x50, 'H'}, 16},                                                // malformed
+		{[]byte{0x14, 'X', 0, 0}, 16},                                          // zero offset
+	}
+	for _, b := range blocks {
+		args := make([]byte, 64)
+		if len(b.src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&b.src[0]))
+		}
+		PutUint64(args, 8, uint64(len(b.src)))
+		PutUint64(args, 16, uint64(cap(b.src)))
+		dst := make([]byte, b.dstSize)
+		if len(dst) > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+		}
+		PutUint64(args, 32, uint64(len(dst)))
+		PutUint64(args, 40, uint64(cap(dst)))
+		argSets = append(argSets, args)
+	}
+
+	// Result offsets: n+48 and code+56.
+	paths, err := k.ProfilePaths("decodeBlockAVX2", argSets, []int{48, 56})
+	if err != nil {
+		t.Fatalf("ProfilePaths: %v", err)
+	}
+
+	// We expect at least 3 distinct paths: success (various n), malformed, zero offset.
+	if len(paths) < 3 {
+		t.Errorf("expected at least 3 distinct paths, got %d", len(paths))
+	}
+	t.Logf("decodeBlockAVX2: %d distinct output paths from %d inputs", len(paths), len(argSets))
+}

verify/differential_test.go

@@ -0,0 +1,295 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"math/rand"
+	"testing"
+	"unsafe"
+)
+
+// decodeBlockGo is a minimal portable LZ4 block decoder used as the
+// differential-testing oracle.  It mirrors the contract of
+// go-lz4's decodeBlockGo: (bytesWritten, code) where code is
+// 0 = ok, 1 = malformed, 2 = zero offset.
+func decodeBlockGo(src, dst []byte) (int, int) {
+	if len(src) == 0 {
+		return 0, 1
+	}
+	si, di := 0, 0
+	for {
+		if si >= len(src) {
+			return 0, 1 // truncated: no token
+		}
+		token := int(src[si])
+		si++
+
+		// Literals.
+		lLen := token >> 4
+		if lLen == 15 {
+			for {
+				if si >= len(src) {
+					return 0, 1
+				}
+				b := int(src[si])
+				si++
+				lLen += b
+				if b != 255 {
+					break
+				}
+			}
+		}
+		if si+lLen > len(src) {
+			return 0, 1 // truncated literals
+		}
+		if di+lLen > len(dst) {
+			return 0, 1 // destination overflow
+		}
+		copy(dst[di:di+lLen], src[si:si+lLen])
+		di += lLen
+		si += lLen
+
+		// End of block.
+		if si >= len(src) {
+			return di, 0
+		}
+
+		// Match offset.
+		if si+2 > len(src) {
+			return 0, 1
+		}
+		offset := int(src[si]) | int(src[si+1])<<8
+		si += 2
+		if offset == 0 {
+			return 0, 2
+		}
+
+		// Match length.
+		mLen := token & 15
+		if mLen == 15 {
+			for {
+				if si >= len(src) {
+					return 0, 1
+				}
+				b := int(src[si])
+				si++
+				mLen += b
+				if b != 255 {
+					break
+				}
+			}
+		}
+		mLen += 4
+
+		// Copy match (overlapping-safe).
+		if di-offset < 0 {
+			return 0, 1 // offset reaches before dst start
+		}
+		if di+mLen > len(dst) {
+			return 0, 1 // destination overflow
+		}
+		for i := 0; i < mLen; i++ {
+			dst[di+i] = dst[di-offset+i]
+		}
+		di += mLen
+	}
+}
+
+// genLZ4Block generates a random valid LZ4 block that decompresses into
+// approximately wantSize bytes.  The block is always well-formed (ends with
+// a literals-only sequence).
+func genLZ4Block(rng *rand.Rand, wantSize int) []byte {
+	var block []byte
+	produced := 0
+	for produced < wantSize {
+		remaining := wantSize - produced
+
+		// Decide: emit a literals+match sequence or the final literals.
+		if remaining <= 8 || rng.Intn(4) == 0 {
+			// Final literals-only sequence.
+			lLen := remaining
+			if lLen > 60 {
+				lLen = 1 + rng.Intn(60)
+			}
+			block = appendToken(block, lLen, 0)
+			for i := 0; i < lLen; i++ {
+				block = append(block, byte(rng.Intn(256)))
+			}
+			produced += lLen
+			break
+		}
+
+		// Literals + match.
+		lLen := rng.Intn(min(16, remaining))
+		if produced+lLen == 0 {
+			lLen = 1 // must have at least 1 literal before the first match
+		}
+		mLenRaw := rng.Intn(12) // match length = mLenRaw + 4
+		mLen := mLenRaw + 4
+		if produced+mLen > remaining {
+			mLen = remaining - produced
+			if mLen < 4 {
+				// Not enough room for a match; emit final literals.
+				lLen = remaining
+				block = appendToken(block, lLen, 0)
+				for i := 0; i < lLen; i++ {
+					block = append(block, byte(rng.Intn(256)))
+				}
+				break
+			}
+			mLenRaw = mLen - 4
+		}
+
+		block = appendToken(block, lLen, mLenRaw)
+		for i := 0; i < lLen; i++ {
+			block = append(block, byte(rng.Intn(256)))
+		}
+		produced += lLen
+
+		// Offset: must be <= produced (can't reference before start).
+		maxOff := produced
+		if maxOff > 65535 {
+			maxOff = 65535
+		}
+		offset := 1 + rng.Intn(maxOff)
+		block = append(block, byte(offset), byte(offset>>8))
+		produced += mLen
+	}
+	return block
+}
+
+// appendToken appends a token (and extension bytes if needed) for the given
+// literal and match lengths.
+func appendToken(block []byte, lLen, mLenRaw int) []byte {
+	lit4 := lLen
+	if lit4 > 15 {
+		lit4 = 15
+	}
+	ml4 := mLenRaw
+	if ml4 > 15 {
+		ml4 = 15
+	}
+	block = append(block, byte(lit4<<4|ml4))
+	// Literal extension bytes.
+	rem := lLen - 15
+	for rem >= 255 {
+		block = append(block, 255)
+		rem -= 255
+	}
+	if lLen >= 15 {
+		block = append(block, byte(rem))
+	}
+	// Match extension bytes.
+	rem = mLenRaw - 15
+	for rem >= 255 {
+		block = append(block, 255)
+		rem -= 255
+	}
+	if mLenRaw >= 15 {
+		block = append(block, byte(rem))
+	}
+	return block
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+// TestDifferentialLZ4Fuzz drives the JIT-assembled decodeBlockAVX2 with
+// random valid LZ4 blocks and compares the output bit-for-bit against the
+// portable Go reference.
+func TestDifferentialLZ4Fuzz(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	const iterations = 5000
+	rng := rand.New(rand.NewSource(42))
+
+	for i := 0; i < iterations; i++ {
+		wantSize := 1 + rng.Intn(4096)
+		src := genLZ4Block(rng, wantSize)
+		dstSize := wantSize + 64 // generous destination
+
+		// Go reference.
+		goDst := make([]byte, dstSize)
+		goN, goCode := decodeBlockGo(src, goDst)
+
+		// JIT kernel.
+		jitDst := make([]byte, dstSize)
+		jitN, jitCode := callDecodeBlockAVX2(t, k, src, jitDst)
+
+		if jitCode != goCode {
+			t.Fatalf("iter %d: code mismatch: JIT=%d, Go=%d (src len=%d)",
+				i, jitCode, goCode, len(src))
+		}
+		if jitCode != 0 {
+			continue // both agree it's malformed/zero-offset
+		}
+		if jitN != goN {
+			t.Fatalf("iter %d: n mismatch: JIT=%d, Go=%d (src len=%d)",
+				i, jitN, goN, len(src))
+		}
+		if !bytes.Equal(jitDst[:jitN], goDst[:goN]) {
+			t.Fatalf("iter %d: output mismatch (n=%d, src len=%d)", i, jitN, len(src))
+		}
+	}
+}
+
+// TestDifferentialLZ4Hostile drives the kernel with random garbage to check
+// that error codes agree with the Go reference (no crashes, same classification).
+func TestDifferentialLZ4Hostile(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	const iterations = 2000
+	rng := rand.New(rand.NewSource(99))
+
+	for i := 0; i < iterations; i++ {
+		srcLen := rng.Intn(128)
+		src := make([]byte, srcLen)
+		rng.Read(src)
+		dstSize := rng.Intn(512)
+		dst := make([]byte, dstSize)
+
+		// Go reference.
+		goDst := make([]byte, dstSize)
+		copy(goDst, dst)
+		_, goCode := decodeBlockGo(src, goDst)
+
+		// JIT kernel.
+		jitDst := make([]byte, dstSize)
+		copy(jitDst, dst)
+		_, jitCode := callDecodeBlockAVX2(t, k, src, jitDst)
+
+		if jitCode != goCode {
+			t.Fatalf("iter %d: hostile code mismatch: JIT=%d, Go=%d (srcLen=%d, dstSize=%d)",
+				i, jitCode, goCode, srcLen, dstSize)
+		}
+	}
+}
+
+// callDecodeBlockAVX2Raw is like callDecodeBlockAVX2 but accepts explicit
+// dst size (for hostile tests where dst may be smaller than the output).
+func callDecodeBlockAVX2Raw(t *testing.T, k *Kernel, src, dst []byte) (int, int) {
+	t.Helper()
+	args := make([]byte, 64)
+	if len(src) > 0 {
+		PutPtr(args, 0, unsafe.Pointer(&src[0]))
+	}
+	PutUint64(args, 8, uint64(len(src)))
+	PutUint64(args, 16, uint64(cap(src)))
+	if len(dst) > 0 {
+		PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+	}
+	PutUint64(args, 32, uint64(len(dst)))
+	PutUint64(args, 40, uint64(cap(dst)))
+
+	out, err := k.CallFunc("decodeBlockAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFunc(decodeBlockAVX2): %v", err)
+	}
+	return int(GetUint64(out, 48)), int(GetUint64(out, 56))
+}

verify/flac_test.go

@@ -0,0 +1,479 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"math/rand"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+const flacKernelPath = "../../go-libraries/go-flac/avx2_amd64.s"
+
+func loadFLACKernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(flacKernelPath); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(flacKernelPath)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", flacKernelPath, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// --- Portable Go references (from go-flac/simd.go) ---
+
+func decodeMono16Go(src []byte, dst []int32) {
+	for i := 0; i < len(dst); i++ {
+		dst[i] = int32(int16(uint16(src[2*i]) | uint16(src[2*i+1])<<8))
+	}
+}
+
+func pack16Go(dst []byte, src []int32) {
+	for i, v := range src {
+		dst[2*i] = byte(v)
+		dst[2*i+1] = byte(v >> 8)
+	}
+}
+
+func decorrelateLeftSideGo(left, right, out []int32) {
+	for i := range left {
+		l := left[i]
+		out[2*i] = l
+		out[2*i+1] = l - right[i]
+	}
+}
+
+func decorrelateSideRightGo(left, right, out []int32) {
+	for i := range left {
+		side := left[i]
+		rch := right[i]
+		out[2*i] = rch + side
+		out[2*i+1] = rch
+	}
+}
+
+func decorrelateMidSideGo(left, right, out []int32) {
+	for i := range left {
+		mid := left[i]
+		side := right[i]
+		mid2 := mid<<1 | (side & 1)
+		out[2*i] = (mid2 + side) >> 1
+		out[2*i+1] = (mid2 - side) >> 1
+	}
+}
+
+func decorrelateInterleaveGo(left, right, out []int32) {
+	for i := range left {
+		out[2*i] = left[i]
+		out[2*i+1] = right[i]
+	}
+}
+
+func analyzeO1RangeGo(swin []int32, dstP []uint32, hist *[32]uint16) (partSum uint64, overflow bool) {
+	swin = swin[:len(dstP)+1]
+	for j := 0; j+1 < len(swin); j++ {
+		r := swin[j+1] - swin[j]
+		if r == -2147483648 { // math.MinInt32
+			overflow = true
+		}
+		f := uint32(r<<1) ^ uint32(r>>31)
+		dstP[j] = f
+		partSum += uint64(f)
+		bl := 0
+		for v := f; v > 0; v >>= 1 {
+			bl++
+		}
+		if bl > 31 {
+			bl = 31
+		}
+		hist[bl]++
+	}
+	return
+}
+
+func analyzeO2RangeGo(swin []int32, dstP []uint32, hist *[32]uint16) (partSum uint64, overflow bool) {
+	swin = swin[:len(dstP)+2]
+	for j := 0; j+2 < len(swin); j++ {
+		r := swin[j+2] - 2*swin[j+1] + swin[j]
+		if r == -2147483648 {
+			overflow = true
+		}
+		f := uint32(r<<1) ^ uint32(r>>31)
+		dstP[j] = f
+		partSum += uint64(f)
+		bl := 0
+		for v := f; v > 0; v >>= 1 {
+			bl++
+		}
+		if bl > 31 {
+			bl = 31
+		}
+		hist[bl]++
+	}
+	return
+}
+
+func analyzeResRangeGo(swin []int32, dstP []uint32, hist *[32]uint16) (partSum uint64, overflow bool) {
+	for j := 0; j < len(swin); j++ {
+		r := swin[j]
+		if r == -2147483648 {
+			overflow = true
+		}
+		f := uint32(r<<1) ^ uint32(r>>31)
+		dstP[j] = f
+		partSum += uint64(f)
+		bl := 0
+		for v := f; v > 0; v >>= 1 {
+			bl++
+		}
+		if bl > 31 {
+			bl = 31
+		}
+		hist[bl]++
+	}
+	return
+}
+
+func decodeMono24Go(src []byte, dst []int32) {
+	for i := 0; i < len(dst); i++ {
+		off := 3 * i
+		u := uint32(src[off]) | uint32(src[off+1])<<8 | uint32(src[off+2])<<16
+		dst[i] = int32(u<<8) >> 8
+	}
+}
+
+// --- Differential tests ---
+
+func TestFLACDecodeMono16(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(7))
+
+	for iter := 0; iter < 500; iter++ {
+		n := rng.Intn(256)
+		src := make([]byte, 2*n)
+		rng.Read(src)
+
+		goDst := make([]int32, n)
+		decodeMono16Go(src, goDst)
+
+		jitDst := make([]int32, n)
+		args := make([]byte, 48)
+		if len(src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(len(src)))
+		PutUint64(args, 16, uint64(cap(src)))
+		if n > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(jitDst)))
+
+		_, err := k.CallFunc("decodeMono16AVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		for i := range goDst {
+			if jitDst[i] != goDst[i] {
+				t.Fatalf("iter %d: mismatch at [%d]: JIT=%d Go=%d", iter, i, jitDst[i], goDst[i])
+			}
+		}
+	}
+}
+
+func TestFLACPack16(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(13))
+
+	for iter := 0; iter < 500; iter++ {
+		n := rng.Intn(256)
+		src := make([]int32, n)
+		for i := range src {
+			src[i] = int32(rng.Intn(65536) - 32768)
+		}
+
+		goDst := make([]byte, 2*n)
+		pack16Go(goDst, src)
+
+		jitDst := make([]byte, 2*n)
+		args := make([]byte, 48)
+		if len(jitDst) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 8, uint64(len(jitDst)))
+		PutUint64(args, 16, uint64(cap(jitDst)))
+		if n > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(src)))
+
+		_, err := k.CallFunc("pack16AVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		if !bytes.Equal(jitDst, goDst) {
+			t.Fatalf("iter %d: output mismatch (n=%d)", iter, n)
+		}
+	}
+}
+
+func TestFLACDecorrelate(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(21))
+
+	kernels := []struct {
+		name string
+		ref  func(left, right, out []int32)
+	}{
+		{"decorrelateLeftSideAVX2", decorrelateLeftSideGo},
+		{"decorrelateSideRightAVX2", decorrelateSideRightGo},
+		{"decorrelateMidSideAVX2", decorrelateMidSideGo},
+		{"decorrelateInterleaveAVX2", decorrelateInterleaveGo},
+	}
+
+	for _, kk := range kernels {
+		t.Run(kk.name, func(t *testing.T) {
+			for iter := 0; iter < 200; iter++ {
+				n := rng.Intn(128)
+				left := make([]int32, n)
+				right := make([]int32, n)
+				for i := range left {
+					left[i] = int32(rng.Intn(1<<24) - 1<<23)
+					right[i] = int32(rng.Intn(1<<24) - 1<<23)
+				}
+
+				goOut := make([]int32, 2*n)
+				kk.ref(left, right, goOut)
+
+				jitOut := make([]int32, 2*n)
+				args := make([]byte, 72)
+				if n > 0 {
+					PutPtr(args, 0, unsafe.Pointer(&left[0]))
+					PutPtr(args, 24, unsafe.Pointer(&right[0]))
+					PutPtr(args, 48, unsafe.Pointer(&jitOut[0]))
+				}
+				PutUint64(args, 8, uint64(n))
+				PutUint64(args, 16, uint64(cap(left)))
+				PutUint64(args, 32, uint64(n))
+				PutUint64(args, 40, uint64(cap(right)))
+				PutUint64(args, 56, uint64(2*n))
+				PutUint64(args, 64, uint64(cap(jitOut)))
+
+				_, err := k.CallFunc(kk.name, args)
+				if err != nil {
+					t.Fatalf("iter %d: %v", iter, err)
+				}
+				for i := range goOut {
+					if jitOut[i] != goOut[i] {
+						t.Fatalf("iter %d: mismatch at [%d]: JIT=%d Go=%d", iter, i, jitOut[i], goOut[i])
+					}
+				}
+			}
+		})
+	}
+}
+
+func TestFLACAnalyzeO1Range(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(33))
+
+	for iter := 0; iter < 300; iter++ {
+		n := 1 + rng.Intn(128) // partition size
+		swin := make([]int32, n+1)
+		for i := range swin {
+			swin[i] = int32(rng.Intn(1<<20) - 1<<19)
+		}
+
+		goDstP := make([]uint32, n)
+		var goHist [32]uint16
+		goSum, goOvf := analyzeO1RangeGo(swin, goDstP, &goHist)
+
+		jitDstP := make([]uint32, n)
+		var jitHist [32]uint16
+		args := make([]byte, 72) // 65 rounded up
+		PutPtr(args, 0, unsafe.Pointer(&swin[0]))
+		PutUint64(args, 8, uint64(len(swin)))
+		PutUint64(args, 16, uint64(cap(swin)))
+		PutPtr(args, 24, unsafe.Pointer(&jitDstP[0]))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(jitDstP)))
+		PutPtr(args, 48, unsafe.Pointer(&jitHist[0]))
+
+		out, err := k.CallFunc("analyzeO1RangeAVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		jitSum := GetUint64(out, 56)
+		jitOvf := out[64] != 0
+
+		if jitSum != goSum {
+			t.Fatalf("iter %d: partSum mismatch: JIT=%d Go=%d", iter, jitSum, goSum)
+		}
+		if jitOvf != goOvf {
+			t.Fatalf("iter %d: overflow mismatch: JIT=%v Go=%v", iter, jitOvf, goOvf)
+		}
+		for i := range goDstP {
+			if jitDstP[i] != goDstP[i] {
+				t.Fatalf("iter %d: dstP[%d] mismatch: JIT=%d Go=%d", iter, i, jitDstP[i], goDstP[i])
+			}
+		}
+		if jitHist != goHist {
+			t.Fatalf("iter %d: hist mismatch: JIT=%v Go=%v", iter, jitHist, goHist)
+		}
+	}
+}
+
+func TestFLACFastStereoSums(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(44))
+
+	for iter := 0; iter < 300; iter++ {
+		n := 1 + rng.Intn(256)
+		left := make([]int32, n)
+		right := make([]int32, n)
+		for i := range left {
+			left[i] = int32(rng.Intn(1<<24) - 1<<23)
+			right[i] = int32(rng.Intn(1<<24) - 1<<23)
+		}
+
+		// Go reference: compute the four sums.
+		var goSums [4]uint64
+		for i := 0; i < n; i++ {
+			l := left[i]
+			r := right[i]
+			side := l - r
+			mid := (l + r) >> 1
+			goSums[0] += foldAbs(l) + foldAbs(r)
+			goSums[1] += foldAbs(l) + foldAbs(side)
+			goSums[2] += foldAbs(side) + foldAbs(r)
+			goSums[3] += foldAbs(mid) + foldAbs(side)
+		}
+
+		var jitSums [4]uint64
+		args := make([]byte, 56)
+		PutPtr(args, 0, unsafe.Pointer(&left[0]))
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(cap(left)))
+		PutPtr(args, 24, unsafe.Pointer(&right[0]))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(right)))
+		PutPtr(args, 48, unsafe.Pointer(&jitSums[0]))
+
+		_, err := k.CallFunc("fastStereoSumsAVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		if jitSums != goSums {
+			t.Fatalf("iter %d: sums mismatch:\n  JIT=%v\n  Go =%v", iter, jitSums, goSums)
+		}
+	}
+}
+
+func foldAbs(v int32) uint64 {
+	return uint64(uint32(v<<1) ^ uint32(v>>31))
+}
+
+// runAnalyzeTest is the shared harness for the analyzeO*Range family.
+func runAnalyzeTest(t *testing.T, k *Kernel, name string, order int, ref func([]int32, []uint32, *[32]uint16) (uint64, bool)) {
+	t.Helper()
+	rng := rand.New(rand.NewSource(int64(order)*100 + 7))
+	for iter := 0; iter < 200; iter++ {
+		n := 1 + rng.Intn(128)
+		swin := make([]int32, n+order)
+		for i := range swin {
+			swin[i] = int32(rng.Intn(1<<20) - 1<<19)
+		}
+
+		goDstP := make([]uint32, n)
+		var goHist [32]uint16
+		goSum, goOvf := ref(swin, goDstP, &goHist)
+
+		jitDstP := make([]uint32, n)
+		var jitHist [32]uint16
+		args := make([]byte, 72)
+		PutPtr(args, 0, unsafe.Pointer(&swin[0]))
+		PutUint64(args, 8, uint64(len(swin)))
+		PutUint64(args, 16, uint64(cap(swin)))
+		PutPtr(args, 24, unsafe.Pointer(&jitDstP[0]))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(jitDstP)))
+		PutPtr(args, 48, unsafe.Pointer(&jitHist[0]))
+
+		out, err := k.CallFunc(name, args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		jitSum := GetUint64(out, 56)
+		jitOvf := out[64] != 0
+
+		if jitSum != goSum {
+			t.Fatalf("iter %d: partSum: JIT=%d Go=%d", iter, jitSum, goSum)
+		}
+		if jitOvf != goOvf {
+			t.Fatalf("iter %d: overflow: JIT=%v Go=%v", iter, jitOvf, goOvf)
+		}
+		for i := range goDstP {
+			if jitDstP[i] != goDstP[i] {
+				t.Fatalf("iter %d: dstP[%d]: JIT=%d Go=%d", iter, i, jitDstP[i], goDstP[i])
+			}
+		}
+		if jitHist != goHist {
+			t.Fatalf("iter %d: hist mismatch", iter)
+		}
+	}
+}
+
+func TestFLACAnalyzeO2Range(t *testing.T) {
+	k := loadFLACKernel(t)
+	runAnalyzeTest(t, k, "analyzeO2RangeAVX2", 2, analyzeO2RangeGo)
+}
+
+func TestFLACAnalyzeResRange(t *testing.T) {
+	k := loadFLACKernel(t)
+	// analyzeResRange has order 0: swin IS the residual (no prediction).
+	runAnalyzeTest(t, k, "analyzeResRangeAVX2", 0, analyzeResRangeGo)
+}
+
+func TestFLACDecodeMono24(t *testing.T) {
+	k := loadFLACKernel(t)
+	rng := rand.New(rand.NewSource(55))
+
+	for iter := 0; iter < 500; iter++ {
+		n := rng.Intn(256)
+		src := make([]byte, 3*n)
+		rng.Read(src)
+
+		goDst := make([]int32, n)
+		decodeMono24Go(src, goDst)
+
+		jitDst := make([]int32, n)
+		args := make([]byte, 48)
+		if len(src) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(len(src)))
+		PutUint64(args, 16, uint64(cap(src)))
+		if n > 0 {
+			PutPtr(args, 24, unsafe.Pointer(&jitDst[0]))
+		}
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(cap(jitDst)))
+
+		_, err := k.CallFunc("decodeMono24AVX2", args)
+		if err != nil {
+			t.Fatalf("iter %d: %v", iter, err)
+		}
+		for i := range goDst {
+			if jitDst[i] != goDst[i] {
+				t.Fatalf("iter %d: dst[%d]: JIT=%d Go=%d", iter, i, jitDst[i], goDst[i])
+			}
+		}
+	}
+}

verify/jit.go

@@ -0,0 +1,88 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Package verify provides the dynamic-analysis substrate for gasm: it
+// JIT-assembles Plan 9 amd64 kernels into executable memory and calls them
+// directly, enabling differential testing against portable Go references,
+// runtime ABI checks and basic-block coverage profiling.
+//
+// The execution model is pure Go (stdlib only): machine code is mapped with
+// syscall.Mmap and invoked through an assembly trampoline that switches to a
+// prepared ABI0 stack.  No cgo, no external toolchain.
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+// Executable maps a copy of code into a read-execute memory region suitable
+// for direct invocation.  The mapping is anonymous and private; the original
+// slice is not retained.  Call Unmap to release the region.
+type Executable struct {
+	addr uintptr // base address of the mapping
+	size int
+	mem  []byte // the mmap'd slice (for Unmap)
+}
+
+// Map copies code into a freshly allocated RX region and returns it.
+// The mapping is PROT_READ|PROT_EXEC; writes are not permitted after the
+// copy, matching W^X policy.
+func Map(code []byte) (*Executable, error) {
+	size := len(code)
+	if size == 0 {
+		return nil, fmt.Errorf("verify: cannot map zero-length code")
+	}
+	// Round up to the page size.
+	const pageSize = 4096
+	mapSize := (size + pageSize - 1) &^ (pageSize - 1)
+
+	mem, err := syscall.Mmap(-1, 0, mapSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, fmt.Errorf("verify: mmap: %w", err)
+	}
+	copy(mem, code)
+
+	// Remove write permission (W^X).
+	if err := syscall.Mprotect(mem, syscall.PROT_READ|syscall.PROT_EXEC); err != nil {
+		syscall.Munmap(mem)
+		return nil, fmt.Errorf("verify: mprotect: %w", err)
+	}
+	return &Executable{
+		addr: uintptr(unsafe.Pointer(&mem[0])),
+		size: size,
+		mem:  mem,
+	}, nil
+}
+
+// Unmap releases the executable region.
+func (e *Executable) Unmap() {
+	if e.mem != nil {
+		syscall.Munmap(e.mem)
+		e.mem = nil
+	}
+}
+
+// FuncAddr returns the absolute address of a function at the given offset
+// within the mapped image.
+func (e *Executable) FuncAddr(offset int) uintptr {
+	return e.addr + uintptr(offset)
+}
+
+// PutUint64 writes v into buf at byte offset off (little-endian).
+func PutUint64(buf []byte, off int, v uint64) {
+	binary.LittleEndian.PutUint64(buf[off:off+8], v)
+}
+
+// GetUint64 reads a little-endian uint64 from buf at byte offset off.
+func GetUint64(buf []byte, off int) uint64 {
+	return binary.LittleEndian.Uint64(buf[off : off+8])
+}
+
+// PutPtr writes a pointer value into buf at byte offset off.
+func PutPtr(buf []byte, off int, p unsafe.Pointer) {
+	binary.LittleEndian.PutUint64(buf[off:off+8], uint64(uintptr(p)))
+}

verify/jit_test.go

@@ -0,0 +1,215 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"testing"
+	"unsafe"
+)
+
+func loadBasic(t *testing.T) *Kernel {
+	t.Helper()
+	k, err := Load("../testdata/verify/basic_amd64.s")
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestJITAdd(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		a, b, want int64
+	}{
+		{0, 0, 0},
+		{1, 2, 3},
+		{-1, 1, 0},
+		{1 << 62, 1 << 62, -9223372036854775808}, // overflow wraps (MinInt64)
+		{-100, -200, -300},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 24)
+		PutUint64(args, 0, uint64(tt.a))
+		PutUint64(args, 8, uint64(tt.b))
+
+		out, err := k.CallFunc("add", args)
+		if err != nil {
+			t.Fatalf("CallFunc(add, %d, %d): %v", tt.a, tt.b, err)
+		}
+		got := int64(GetUint64(out, 16))
+		if got != tt.want {
+			t.Errorf("add(%d, %d) = %d, want %d", tt.a, tt.b, got, tt.want)
+		}
+	}
+}
+
+func TestJITSum(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		data []int64
+		want int64
+	}{
+		{nil, 0},
+		{[]int64{1}, 1},
+		{[]int64{1, 2, 3, 4, 5}, 15},
+		{[]int64{-10, 20, -30, 40}, 20},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 32)
+		if len(tt.data) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&tt.data[0]))
+		}
+		PutUint64(args, 8, uint64(len(tt.data)))
+		PutUint64(args, 16, uint64(cap(tt.data)))
+
+		out, err := k.CallFunc("sum", args)
+		if err != nil {
+			t.Fatalf("CallFunc(sum, %v): %v", tt.data, err)
+		}
+		got := int64(GetUint64(out, 24))
+		if got != tt.want {
+			t.Errorf("sum(%v) = %d, want %d", tt.data, got, tt.want)
+		}
+	}
+}
+
+func TestJITWideCopy(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		name string
+		n    int
+	}{
+		{"empty", 0},
+		{"tiny", 7},
+		{"exact32", 32},
+		{"overlap_range", 48},
+		{"exact64", 64},
+		{"unaligned", 45},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			src := make([]byte, tt.n)
+			for i := range src {
+				src[i] = byte(i * 7)
+			}
+			dst := make([]byte, tt.n)
+
+			args := make([]byte, 48)
+			if tt.n > 0 {
+				PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+				PutPtr(args, 24, unsafe.Pointer(&src[0]))
+			}
+			PutUint64(args, 8, uint64(tt.n))  // dst_len
+			PutUint64(args, 16, uint64(tt.n)) // dst_cap
+			PutUint64(args, 32, uint64(tt.n)) // src_len
+			PutUint64(args, 40, uint64(tt.n)) // src_cap
+
+			_, err := k.CallFunc("wideCopy", args)
+			if err != nil {
+				t.Fatalf("CallFunc(wideCopy): %v", err)
+			}
+			if !bytes.Equal(dst, src) {
+				t.Errorf("wideCopy: dst ≠ src\n  got  %x\n  want %x", dst, src)
+			}
+		})
+	}
+}
+
+func TestKernelFuncNames(t *testing.T) {
+	k := loadBasic(t)
+	names := k.FuncNames()
+	want := []string{"add", "sum", "wideCopy"}
+	if len(names) != len(want) {
+		t.Fatalf("FuncNames() = %v, want %v", names, want)
+	}
+	for i, n := range names {
+		if n != want[i] {
+			t.Errorf("FuncNames()[%d] = %q, want %q", i, n, want[i])
+		}
+	}
+}
+
+func TestKernelFuncNotFound(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("nonexistent", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+}
+
+func TestKernelArgTooSmall(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("add", make([]byte, 8)) // needs 24
+	if err == nil {
+		t.Fatal("expected error for too-small arg block")
+	}
+}
+
+func TestMapZeroLength(t *testing.T) {
+	_, err := Map(nil)
+	if err == nil {
+		t.Fatal("expected error for zero-length code")
+	}
+}
+
+func TestLoadSourceError(t *testing.T) {
+	_, err := LoadSource("bad.s", "TEXT ·f(SB), NOSPLIT\n\tBADINSTRUCTION\n")
+	// The parser may or may not error on unknown instructions (it's
+	// error-tolerant), but the assembler will reject it.
+	if err == nil {
+		t.Log("LoadSource succeeded unexpectedly (parser is error-tolerant)")
+	}
+}
+
+func TestLoadSourceParseError(t *testing.T) {
+	// A completely invalid file that the parser rejects.
+	_, err := LoadSource("empty.s", "")
+	if err != nil {
+		t.Logf("expected: %v", err)
+	}
+}
+
+func TestFuncLookup(t *testing.T) {
+	k := loadBasic(t)
+	fl, err := k.Func("add")
+	if err != nil {
+		t.Fatalf("Func(add): %v", err)
+	}
+	if fl.Name != "add" {
+		t.Errorf("Func(add).Name = %q, want %q", fl.Name, "add")
+	}
+	if fl.Args != 24 {
+		t.Errorf("Func(add).Args = %d, want 24", fl.Args)
+	}
+	_, err = k.Func("nonexistent")
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+}
+
+func TestABIReportString(t *testing.T) {
+	r := ABIReport{}
+	if r.String() != "ABI clean" {
+		t.Errorf("clean report = %q", r.String())
+	}
+	r.BPClobbered = true
+	if r.OK() {
+		t.Error("expected not OK with BP clobbered")
+	}
+	s := r.String()
+	if s == "ABI clean" {
+		t.Error("expected violation string, got clean")
+	}
+	r.R14Clobbered = true
+	r.RedZoneHit = true
+	s = r.String()
+	if s == "ABI clean" {
+		t.Error("expected violation string for all flags")
+	}
+}

verify/lz4_test.go

@@ -0,0 +1,160 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+// lz4KernelPath is the sibling repository's AVX2 kernel, used for
+// integration testing.  The test is skipped when the file is absent
+// (e.g. in CI without the sibling checkout).
+const lz4KernelPath = "../../go-libraries/go-lz4/avx2_amd64.s"
+
+func loadLZ4Kernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(lz4KernelPath); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(lz4KernelPath)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", lz4KernelPath, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// callDecodeBlockAVX2 invokes the JIT-assembled decodeBlockAVX2 with the
+// given src and dst buffers, returning (n, code).
+func callDecodeBlockAVX2(t *testing.T, k *Kernel, src, dst []byte) (int, int) {
+	t.Helper()
+	args := make([]byte, 64)
+	if len(src) > 0 {
+		PutPtr(args, 0, unsafe.Pointer(&src[0]))
+	}
+	PutUint64(args, 8, uint64(len(src)))
+	PutUint64(args, 16, uint64(cap(src)))
+	if len(dst) > 0 {
+		PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+	}
+	PutUint64(args, 32, uint64(len(dst)))
+	PutUint64(args, 40, uint64(cap(dst)))
+
+	out, err := k.CallFunc("decodeBlockAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFunc(decodeBlockAVX2): %v", err)
+	}
+	return int(GetUint64(out, 48)), int(GetUint64(out, 56))
+}
+
+func TestLZ4DecodeKnownAnswers(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	tests := []struct {
+		name     string
+		src      []byte
+		dstSize  int
+		wantDst  []byte
+		wantN    int
+		wantCode int
+	}{
+		{
+			name:     "literals_only",
+			src:      []byte{0x50, 'H', 'e', 'l', 'l', 'o'},
+			dstSize:  16,
+			wantDst:  []byte("Hello"),
+			wantN:    5,
+			wantCode: 0,
+		},
+		{
+			name:     "literals_and_match",
+			src:      []byte{0x54, 'A', 'A', 'A', 'A', 'A', 0x05, 0x00, 0x30, 'B', 'B', 'B'},
+			dstSize:  32,
+			wantDst:  []byte("AAAAAAAAAAAAABBB"),
+			wantN:    16,
+			wantCode: 0,
+		},
+		{
+			name: "overlapping_match",
+			// 1 literal 'X', then match offset=1 length=4+4=8 → "XXXXXXXXX",
+			// then final 1 literal 'Y'.
+			src:      []byte{0x14, 'X', 0x01, 0x00, 0x10, 'Y'},
+			dstSize:  16,
+			wantDst:  []byte("XXXXXXXXXY"),
+			wantN:    10,
+			wantCode: 0,
+		},
+		{
+			name:     "malformed_truncated",
+			src:      []byte{0x50, 'H', 'e'}, // claims 5 literals, has 2
+			dstSize:  16,
+			wantN:    0,
+			wantCode: 1,
+		},
+		{
+			name:     "zero_offset",
+			src:      []byte{0x14, 'X', 0x00, 0x00},
+			dstSize:  16,
+			wantN:    0,
+			wantCode: 2,
+		},
+		{
+			name:     "empty_token",
+			src:      []byte{0x00}, // 0 literals, end of block
+			dstSize:  16,
+			wantDst:  nil,
+			wantN:    0,
+			wantCode: 0,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dst := make([]byte, tt.dstSize)
+			n, code := callDecodeBlockAVX2(t, k, tt.src, dst)
+			if n != tt.wantN || code != tt.wantCode {
+				t.Fatalf("decodeBlockAVX2: got (n=%d, code=%d), want (n=%d, code=%d)",
+					n, code, tt.wantN, tt.wantCode)
+			}
+			if tt.wantCode == 0 && tt.wantDst != nil {
+				if !bytes.Equal(dst[:n], tt.wantDst) {
+					t.Errorf("output mismatch:\n  got  %q\n  want %q", dst[:n], tt.wantDst)
+				}
+			}
+		})
+	}
+}
+
+func TestLZ4WideCopyAVX2(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	sizes := []int{0, 1, 15, 16, 31, 32, 33, 63, 64, 100, 256, 1024}
+	for _, n := range sizes {
+		src := make([]byte, n)
+		for i := range src {
+			src[i] = byte(i*13 + 7)
+		}
+		dst := make([]byte, n)
+
+		args := make([]byte, 48)
+		if n > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(n))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(n))
+
+		_, err := k.CallFunc("wideCopyAVX2", args)
+		if err != nil {
+			t.Fatalf("wideCopyAVX2(n=%d): %v", n, err)
+		}
+		if !bytes.Equal(dst, src) {
+			t.Errorf("wideCopyAVX2(n=%d): output mismatch", n)
+		}
+	}
+}

verify/trampoline_amd64.s

@@ -0,0 +1,30 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// ABI0 JIT trampoline.  enterJIT switches from the Go stack to a prepared
+// stack and jumps to the assembled function; when the function RETs, control
+// lands in leaveJIT, which restores the Go stack and returns to the Go caller.
+//
+// The prepared stack must begin with the address of leaveJIT (the return
+// address the JIT function will pop), followed by the function's ABI0
+// argument area.
+//
+// Single-threaded: savedSP is a package global, so only one JIT call may be
+// in flight at a time.  gasm verify runs sequentially.
+
+// func enterJIT(fn uintptr, stack uintptr)
+// Switches to the prepared stack and jumps to fn.  Does not return normally;
+// the JIT function's RET transfers control to leaveJIT.
+TEXT ·enterJIT(SB), NOSPLIT, $0-16
+	MOVQ fn+0(FP), AX      // target function address (before SP switch)
+	MOVQ SP, ·savedSP(SB)  // preserve the Go stack pointer
+	MOVQ stack+8(FP), SP   // switch to the prepared stack
+	JMP  AX
+
+// func leaveJIT()
+// Restores the Go stack pointer and returns to enterJIT's caller.
+TEXT ·leaveJIT(SB), NOSPLIT, $0-0
+	MOVQ ·savedSP(SB), SP
+	RET

verify/verify.go

@@ -0,0 +1,122 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"fmt"
+	"os"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// Kernel is a JIT-loaded assembly image ready for direct invocation.
+// It wraps an executable memory mapping and the function layout metadata
+// needed to marshal ABI0 calls.
+type Kernel struct {
+	exec  *Executable
+	img   *asm.Image
+	funcs map[string]int // function name → index into img.Funcs
+}
+
+// Load parses, assembles and maps a .s file into executable memory.
+// The returned Kernel is ready for Call.  The caller must call Close to
+// release the mapping.
+func Load(path string) (*Kernel, error) {
+	src, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("verify: %w", err)
+	}
+	return LoadSource(path, string(src))
+}
+
+// LoadSource parses, assembles and maps assembly source into executable memory.
+func LoadSource(filename, src string) (*Kernel, error) {
+	file, errs := parser.Parse(filename, src)
+	if len(errs) > 0 {
+		return nil, fmt.Errorf("verify: parse %s: %v", filename, errs[0])
+	}
+	return LoadAST(file)
+}
+
+// LoadAST assembles a parsed AST file and maps the result into executable
+// memory.
+func LoadAST(file *ast.File) (*Kernel, error) {
+	img, err := asm.AssembleFile(file)
+	if err != nil {
+		return nil, fmt.Errorf("verify: assemble: %w", err)
+	}
+	if len(img.Externals) > 0 {
+		return nil, fmt.Errorf("verify: unresolved external symbols: %v", img.Externals)
+	}
+	code := img.Bytes()
+	exec, err := Map(code)
+	if err != nil {
+		return nil, err
+	}
+	funcs := make(map[string]int, len(img.Funcs))
+	for i, f := range img.Funcs {
+		funcs[f.Name] = i
+	}
+	return &Kernel{exec: exec, img: img, funcs: funcs}, nil
+}
+
+// Func returns the layout metadata for the named function.
+func (k *Kernel) Func(name string) (asm.FuncLayout, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return asm.FuncLayout{}, fmt.Errorf("verify: function %q not found", name)
+	}
+	return k.img.Funcs[idx], nil
+}
+
+// FuncNames returns the names of all functions in the kernel, in source order.
+func (k *Kernel) FuncNames() []string {
+	names := make([]string, len(k.img.Funcs))
+	for i, f := range k.img.Funcs {
+		names[i] = f.Name
+	}
+	return names
+}
+
+// CallFunc invokes the named function with the given ABI0 argument block.
+// The arg block is the raw bytes of the function's argument/result area
+// (as declared by the TEXT $frame-args suffix).  Returns the arg block
+// after the call (with any results written back by the function).
+func (k *Kernel) CallFunc(name string, args []byte) ([]byte, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+	if len(args) < fl.Args {
+		return nil, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
+	}
+	fnAddr := k.exec.FuncAddr(fl.Offset)
+	return Call(fnAddr, args)
+}
+
+// CallFuncChecked invokes the named function with ABI sentinels and a
+// red-zone canary, returning the argument block and an ABIReport that
+// records any callee-saved register or red-zone violations.
+func (k *Kernel) CallFuncChecked(name string, args []byte) ([]byte, ABIReport, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, ABIReport{}, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+	if len(args) < fl.Args {
+		return nil, ABIReport{}, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
+	}
+	fnAddr := k.exec.FuncAddr(fl.Offset)
+	return CallChecked(fnAddr, args)
+}
+
+// Close releases the executable mapping.
+func (k *Kernel) Close() {
+	if k.exec != nil {
+		k.exec.Unmap()
+	}
+}